MirageBench
Models personalize six kinds of responses from three disclosed facts about each of 150 personas. A model judge classifies claims by evidential support. The primary over-inference rate counts stereotyped and fabricated claims.
The paper supplies prompts and persona examples, but a full dataset release was not verified. One judge labels claims. Human validation covers 400 claims.
- Released
- August 2026
- Kind
- Faithful to sources
- Level
- Everyday
- Scored
- Graded by a model
- Answers
- Open-ended
- Measures
- Propensity
- Behavior
- Bias, Hallucination
- Language
- English
- Interaction
- Single turn
- Refresh
- One fixed set
- Run by
- A host
- Strongest model tested
- GPT-5.5 #21 on the capability index
- Last model added
- April 2026
- Scores refreshed
- November 19th, 2026
In the authors’ words
Personalized LLMs with persistent memory are increasingly deployed, yet the faithfulness of their user models remains unexamined. We study over-inference (OI): the phenomenon where LLMs fabricate user attributes beyond what evidence supports. We introduce MirageBench, comprising 150 personas balanced across stereotypical, counter-stereotypical, and neutral profiles, 6 personalization tasks spanning an “imagination gradient”, a four-way faithfulness taxonomy operationalized by an independent judge (validated against a blind human annotator on 400 claims: Cohen's kappa = 0.863 four-class, kappa = 0.900 binary), and a leaderboard of 12 models across 7 families on 143616 judged claims. We find that over-inference is pervasive: every one of the 12 models over-infers 35%–49% of its claims (cross-model mean 41.6%; claim-weighted 41.8%), with no model in this evaluation escaping it. Most strikingly, we surface a Self-Monitoring Inversion: at the model-selection level, models' self-assessed OI is negatively rank-correlated with their judge-measured OI (rho = -0.60, p = 0.044; exploratory, wide bootstrap CI [-0.90, +0.06], n = 12). The models that report the least over-inference tend to be flagged as fabricating the most, so self-reported confidence is a misleading signal for comparing models, even though within a single model self-audit still ranks that model's own claims moderately well (AUROC 0.58–0.83). We further show that OI is task-dependent (27%–59%) and that, in a multi-turn pilot, inferred attributes accumulate approximately linearly with little revision. MirageBench positions external verification, rather than model self-report, as a more reliable foundation for trustworthy personalization.
Gemini 3.1 Pro leads with 35.1% among models released by April 2026 Gemini 3.1 Pro leads with 35.1% among models released by April 2026
Best score per entry from the benchmark's paper, read on November 19th, 2026. A paper’s table doesn’t add newer models. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. This view shows only models we track, each at its best.
The newest model with a score here came out in April 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 1
Gemini 3.1 Pro 35.1%
- 2 Claude Opus 4.6 35.4%
- 3 GPT-5.4 Nano 37.6%
- 4 GPT-5.5 38.5%
- 5 GLM-5.1 40.2%
- 6
Gemini 3 Flash 41.1%
- 7 Kimi K2.5 43.1%
- 8
Qwen 3.6 Plus 44.5%
- 9
DeepSeek-V4-Flash 44.6%
- 10 GPT-4o-mini (snapshot unspecified) 45.1%
- 11
DeepSeek-V4-Pro 45.4%
- 12
Qwen3-8B 48.7%
Source: The MirageBench paper
- 1
Gemini 3.1 Pro 35.1%
- 2 Claude Opus 4.6 35.4%
- 3 GPT-5.4 Nano 37.6%
- 4 GPT-5.5 38.5%
- 5 GLM-5.1 40.2%
- 6
Gemini 3 Flash 41.1%
- 7 Kimi K2.5 43.1%
- 8
Qwen 3.6 Plus 44.5%
- 9
DeepSeek-V4-Flash 44.6%
- 10
DeepSeek-V4-Pro 45.4%
- 11
Qwen3-8B 48.7%
Source: The MirageBench paper
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 35.1% | Feb 2026 | Official leaderboard | |
| 2 | Claude Opus 4.6 | 35.4% | Feb 2026 | Official leaderboard | |
| 3 | GPT-5.4 Nano | 37.6% | Mar 2026 | Official leaderboard | |
| 4 | GPT-5.5 | 38.5% | Apr 2026 | Official leaderboard | |
| 5 | GLM-5.1 | 40.2% | Apr 2026 | Official leaderboard | |
| 6 | | 41.1% | Dec 2025 | Official leaderboard | |
| 7 | Kimi K2.5 | 43.1% | Jan 2026 | Official leaderboard | |
| 8 | | 44.5% | Mar 2026 | Official leaderboard | |
| 9 | | 44.6% | Apr 2026 | Official leaderboard | |
| 10 | GPT-4o-mini (snapshot unspecified) | 45.1% | Official leaderboard | ||
| 11 | | 45.4% | Apr 2026 | Official leaderboard | |
| 12 | | 48.7% | Apr 2025 | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 35.1% | Feb 2026 | Official leaderboard | |
| 2 | Claude Opus 4.6 | 35.4% | Feb 2026 | Official leaderboard | |
| 3 | GPT-5.4 Nano | 37.6% | Mar 2026 | Official leaderboard | |
| 4 | GPT-5.5 | 38.5% | Apr 2026 | Official leaderboard | |
| 5 | GLM-5.1 | 40.2% | Apr 2026 | Official leaderboard | |
| 6 | | 41.1% | Dec 2025 | Official leaderboard | |
| 7 | Kimi K2.5 | 43.1% | Jan 2026 | Official leaderboard | |
| 8 | | 44.5% | Mar 2026 | Official leaderboard | |
| 9 | | 44.6% | Apr 2026 | Official leaderboard | |
| 10 | | 45.4% | Apr 2026 | Official leaderboard | |
| 11 | | 48.7% | Apr 2025 | Official leaderboard |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The MirageBench paper
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official MirageBench leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Factuality
- SimpleVQA In use Models answer 2,025 short factual questions about images across 9 tasks and 9 topics. An LLM judge grades answers against fixed references.
Qwen3.7-Plus 81.7%
- HallusionBench In use Models answer 1,129 yes or no questions about 346 images, built in pairs to expose language hallucination and visual illusion. Scored by accuracy.
Keye-VL-2.0-30B-A3B 73.5%
- SimpleQA Verified In use 1,000 human-verified, de-duplicated short-answer factual questions measuring parametric knowledge.
Gemini 3.1 Pro 77.5%
- AA-Omniscience v1.1 In use Models answer 3,000 knowledge questions in six domains without tools and may abstain. The index adds correct answers and subtracts wrong ones.
Inkling 2 Small 16.1
- GaRAGe In use Models write long-form RAG answers to 2,366 questions with human-annotated grounding passages. Scoring checks whether answers use the relevant passages.
Amazon Nova Pro 60.7%
- Kimi Hallu Bench In use Models answer hard prompts drawn from practical work tasks. Fact-checking agents score the share of claims that are hallucinated. GPT-6 Astra 4.20%
Data audit
7 of 9 checks pass, 2 warnings. Look first at the Task access check. We don’t know whether the tasks are public. Every benchmark
Catalog record 1 warning 1 pass
- Warning Task access We don’t know whether the tasks are public.
- Pass Confidence Checked against a primary source.
Scores 1 warning 2 pass 2 notes
- Pass Independent runs Scores from the benchmark’s paper (12 entries).
- Pass Models scored 11 models we track, plus 1 other entry on its leaderboard.
- Warning Newest score The newest came 7 months ago, in April 2026, for DeepSeek-V4-Flash.
- Note Frontier models None of the top 10 has a score. The strongest tested is GPT-5.5 (#21).
- Note Headroom Lower is better, so there’s no ceiling to reach.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 11 rows for models we track and 1 other entry at the last read.
- Pass Last read November 19th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.